LDA Based Feature Extraction And It Sapplication To Face Recognition

Data dimensionality reduction is a key technology in the process of feature extraction, since the data of face image representation is always high dimension, which usually lead to the curse of dimensionality problem. In order to overcome the above problem, dimensionality reduction of high-dimensional data has become a very important step. In the combination of the data dimension reduction method and face image feature extraction method, the dimension reduction method based on subspace is the most widely used. And the method based on linear discriminant analysis(LDA) has become a hot research topic in recent years. However, since the separable criterion of traditional LDA method is not directly related with the output subspace classification accuracy, therefore, reducing the performance of LDA algorithm. In view of the above problems, the concrete research content of this paper is mainly divided into the following 3 parts:(1)Analyse and compare the common face image dimension reduction methodData dimensionality reduction algorithm can be divided into linear and nonlinear. The linear dimension reduction techniques have been widely applied because of its simple implementation. This paper analyzes two kinds of dimension reduction methods in theory and we hope to combine the characteristics of face image data to analyze advantages and disadvantages of the various dimension reduction methods.(2) A Comparative Study of LDA and Its Extension Methods in Face RecognitionThe best projection direction is obtained by calculating the extremum of the Fisher criterion function in the LDA method. However, the scatter matrix as the foundation of the LDA in high dimension small sample is irreversible(singular), so that the traditional LDA can not be calculated directly. Therefore, the within class scatter matrix and the between class scatter matrix is studied. Research methods have the following kinds: a. through the PCA dimension reduction pretreatment, make the within class scatter matrix after dimensionality reduction is invertible, b. weight the within class scatter matrix and the between class scatter matrix, c. mapping the original data to nonlinear high dimension space by kernel function, and then perform the feature extraction. Through the above research, and hope to be combined with the experimental to analyse the LDA and its extension methods in different face database.(3)Use characteristics of various feature extraction methods for the minimum distance classifier, KNN classifier and SVM classifierDue to the characteristics of different face databases are different, so take a variety ofclassification methods to study them. Use feature matrices of various feature extraction methods for training in the three representative classifier. Hope to compare various classification methods through experiments and find out the cause of different classification performance.Linear discriminant analysis is an important feature extraction method, an improved LDA algorithm is proposed in this paper. This algorithm standardize the the distance between sample class, namely, in the definition of the sample between class scatter only consider the direction of between class distance, while neglecting the distance size, through this method make all categories are treated fairly. Moreover, by redefining the within-class scatter matrices and introducing the normalized parameter to control the bias and variance of eigenvalues, in addition, it makes the between-class scatter matrices weight and could be avoiding the overlapping of neighboring class samples.